CN115238576A - A Vehicle Longitudinal Dynamics Identification Method Combining Neural Network and Physical Model - Google Patents

Abstract

The invention discloses a vehicle longitudinal dynamics identification method combining a neural network and a physical model, and the method comprises the following steps of S1, driving an unmanned vehicle to run, and recording running data comprising gear positions of a gearbox; s2, processing is carried out according to the driving data to obtain a sample data set for training; s3, respectively training the accelerator neural network model and the brake neural network model through the sample data set; and S4, combining the trained accelerator neural network model and the trained braking neural network model with the vehicle driving physical model and the vehicle braking physical model to obtain a longitudinal dynamics model of the unmanned vehicle, so that a physical equation constructed by vehicle gear and road gradient information is combined with the neural network, the longitudinal dynamics of the vehicle can be better represented, and the unmanned vehicle can be suitable for more types of vehicles.

Description

A Vehicle Longitudinal Dynamics Identification Method Combining Neural Network and Physical Model

technical field

The invention relates to the technical field of automatic driving vehicles, in particular to a vehicle longitudinal dynamics identification method combining a neural network and a physical model.

Background technique

Autonomous driving technology has become an important carrier for collaborative innovation in many fields and the construction of a new transportation system. It has great strategic significance in reshaping the industrial ecology, promoting technological innovation, improving traffic safety, and achieving energy conservation and emission reduction. Autonomous driving technology can be divided into key technologies such as positioning, perception, motion prediction, decision planning, motion control, and simulation, among which motion control is one of the core technologies for autonomous vehicles to achieve safe and smooth driving. To realize the motion control of autonomous vehicles, the analysis and modeling of vehicle dynamics is the key to vehicle dynamics simulation and motion control. Vehicles are complex electromechanical-hydraulic devices with multiple subsystems, and their dynamic models are generally constructed based on two ideas: physical models and data-driven models. The process of building a physical model is also modeled in a white box, by analyzing the structure, composition, and laws of motion of the system, applying known laws, theorems, and principles, such as principles of mechanics, laws of biology, Newton's laws, energy balance equations, heat transfer The principle of mass transfer, etc., is deduced by mathematical methods, and the physical model of the system is established. The construction process of data-driven model is also called black-box modeling, which uses data for modeling.

In the prior art: The invention patent with publication number CN202010247999.5 discloses a vehicle longitudinal dynamics system identification method. By collecting the input and output data of the vehicle, the frequency domain response analysis method is used to estimate the throttle opening of the gear. Acceleration characteristics and corresponding mathematical models of the vehicle in each gear are obtained, and the vehicle longitudinal dynamics system identification is completed. However, this method relies on a large amount of human experience for analysis, and large-scale vehicle identification will consume a lot of manpower and financial resources, and the mathematical model cannot cover all working conditions.

The invention patent with publication number CN110967991A discloses a method, device, on-board controller and unmanned vehicle for determining vehicle control parameters, which can automatically identify the longitudinal driving and braking system of the vehicle by relying on a large amount of data only on neural networks, and can It is used to identify the longitudinal dynamics model of large-scale vehicles and quickly determine the parameters of motion control, but this method is mostly used on hard roads such as structured cement and asphalt. The acceleration fluctuation caused by the blocking process cannot be well fitted.

The invention patent publication number CN113657036A discloses a method for realizing vehicle dynamics simulation based on neural network and physical model. The method also combines the vehicle lateral dynamics model with the neural network model to improve the accuracy of vehicle lateral dynamics modeling. . But this method is only used for vehicle lateral dynamics calibration.

SUMMARY OF THE INVENTION

Aiming at the problem that the longitudinal dynamic model of the self-driving vehicle is complex, and has strong nonlinear and time-varying characteristics, the modeling is difficult. Longitudinal dynamics model and inverse dynamics model of the vehicle. Among them, the longitudinal dynamics model can be applied to the longitudinal dynamics simulation modeling of a specific vehicle, and can simulate the longitudinal driving and braking response results close to the real vehicle; the inverse dynamics model can be transformed into a vehicle control parameter map and applied to no The longitudinal speed control of the human-driven vehicle realizes the mapping of the desired acceleration output by the longitudinal speed control algorithm to the accelerator opening and braking opening commands that can be received by the executive layer.

The present invention aims to provide a vehicle longitudinal dynamics identification method combining a neural network and a physical model, so as to solve or improve at least one of the above technical problems.

In view of this, a first aspect of the present invention is to provide a vehicle longitudinal dynamics identification method combining a neural network and a physical model.

A first aspect of the present invention provides a vehicle longitudinal dynamics identification method combining a neural network and a physical model, including: the longitudinal dynamics model is:

a _r =MODEL(v _x , φ _t , φ _b , _ig , θ _{road gradient} ), where: a _r is the actual acceleration of the vehicle, v _x is the longitudinal speed of the vehicle, φ _t is the accelerator opening, and φ _b is the braking Opening, _ig are the transmission reduction ratio, θ is the road gradient, and the longitudinal dynamics model is calibrated by the method, including the following steps: S1, driving the unmanned vehicle to travel, and recording the driving data including the transmission gear; S2, process the driving data to obtain a sample data set for training; S3, train the throttle neural network model and the braking neural network model respectively through the sample data set; S4, train the throttle neural network model, the brake neural network model after training The dynamic neural network model is combined with the vehicle driving physical model and the vehicle braking physical model to obtain the longitudinal dynamics model of the unmanned vehicle; wherein, the driving data includes: accelerator opening, braking opening, longitudinal speed, Longitudinal acceleration and transmission gear.

The invention provides a vehicle longitudinal dynamics identification method combined with a neural network and a physical model, aiming at the problem that the longitudinal speed control of an autonomous driving vehicle faces the complex longitudinal dynamics model, which has strong nonlinear and time-varying characteristics, which leads to difficult modeling. Combining the physical equation constructed by the vehicle's transmission gear and road gradient information with the neural network can better characterize the vehicle's longitudinal dynamics, and can be applied to more types of vehicles. The problem of constructing the overall longitudinal dynamics model of the vehicle is simplified as Only for the modeling problem of the equivalent acceleration generated by the drive system and the braking system, the obtained longitudinal dynamics model is more accurate.

In addition, the technical solutions provided according to the embodiments of the present invention may also have the following additional technical features:

In any of the above technical solutions, through the reverse conversion of the longitudinal dynamics model, the vehicle longitudinal control parameter mapping relationship is obtained, which is specifically the following formula:

(φ _t , φ _b )=MAP(v _x , a _{desired speed} , θ _{road gradient} , i _g )

Wherein, the vehicle longitudinal control parameter mapping relationship is used for the control of the longitudinal speed of the unmanned vehicle, given the current vehicle operating state (v _x , θ _{road gradient} , i _g ) and the value of the longitudinal control rate (a _{desired speed} ) When , query to obtain the accelerator and brake opening values (φ _t , φ _b ) to be executed by the vehicle.

In this technical solution, the inverse dynamics model can be transformed into a vehicle control parameter map and applied to the longitudinal speed control of the unmanned vehicle, so as to realize the desired acceleration output by the longitudinal speed control algorithm to the accelerator opening and braking that can be received by the executive layer. Mapping of opening instructions;

This mapping relationship can be used for vehicle longitudinal speed control. The mapping relationship of specific vehicle longitudinal control parameters means that the current vehicle transmission gear reduction ratio, vehicle longitudinal speed, road gradient, and expected acceleration are used as inputs to obtain the accelerator opening value and brake opening value that the vehicle control-by-wire system can receive. .

In any of the above technical solutions, the step of S1 includes: S11, setting different accelerator opening degrees and different brake opening degrees; S12, driving the unmanned vehicle on a horizontal road at different speeds within a preset range driving; S13, collect the accelerator opening, brake opening, longitudinal speed, longitudinal acceleration, and transmission gear position when the unmanned vehicle is running; wherein, the preset range is 0km/h-30km/h, and the different The accelerator opening and braking opening are discrete opening commands of 0%, 10%, 20%, ..., 100%.

In this technical solution, a flat road is selected, and the autonomous vehicle uses the pre-built acquisition program to conduct the acquisition experiment, and collects the vehicle longitudinal speed, longitudinal acceleration, and transmission gear (including the reduction ratio) according to the fixed accelerator/brake opening requirements. Wait for the vehicle status data to get real vehicle data [accelerator opening, brake opening, longitudinal speed, longitudinal acceleration, transmission gear];

It is mainly used in vehicles with a speed of less than 30km/h, such as mining wide-body dump trucks, muck trucks and other engineering vehicles, so the air resistance can be ignored.

In any of the above technical solutions, the step of S2 includes: S21, estimating the longitudinal speed and longitudinal acceleration of the vehicle; S22, estimating the rolling resistance acceleration term; S23, estimating the acceleration response delay time, and Perform data alignment.

In this technical solution, the data is preprocessed to form a training data set for neural network training. The data processing mainly includes the following steps: estimating the longitudinal speed and longitudinal acceleration of the vehicle, estimating the rolling resistance acceleration term, estimating the Acceleration response delay time estimation.

In any of the above technical solutions, the estimation method of S21 is:

S211, firstly use the following formula to limit and filter the measured longitudinal acceleration, specifically:

Δa=a _mea (k)-a _mea (k-1)

Among them, Δa is the increase value from the (k-1)th sampling to the (k)th sampling, a _mea represents the measured value, a _{mea, limit} is the measured value after limiting;

S212, perform reprocessing to obtain the longitudinal velocity and longitudinal acceleration of the vehicle, specifically:

The discretized state transition equation and observation equation are constructed based on the vehicle kinematics model,

The state transition equation is:

即X(k)＝AX(k-1)+w(k-1)

That is, X(k)=AX(k-1)+w(k-1)

The observation equation is:

即Z(k)＝HX(k)+ε(k)

That is, Z(k)=HX(k)+ε(k)

为状态矢量、

为系统矩阵、

为传感器观测值、

为观测矩阵、T为数据采样周期、k为离散化的采样序列、

为系统的过程噪声、

为传感器观测噪声、T为数据采样周期、k为离散化的采样序列；in,

is the state vector,

is the system matrix,

are the sensor observations,

is the observation matrix, T is the data sampling period, k is the discretized sampling sequence,

is the process noise of the system,

is the sensor observation noise, T is the data sampling period, and k is the discretized sampling sequence;

S213, finally predict and update iteratively according to the following formula, specifically:

为k时刻的最优估计状态矢量、

为k时刻的预测状态矢量、P(k)为k时刻的协方差矩阵更新值、

为k时刻的协方差矩阵预测值、K_G(k)为k时刻的卡尔曼矩阵、

为预测过程误差协方差矩阵、σ_v为速度状态转移方程的过程噪声标准差、σ_a为加速度状态转移方程的过程噪声标准差、

为传感器观测噪声协方差矩阵、σ_v，m为速度传感器噪声的标准差、σ_a，m为加速度传感器噪声的标准差、I为单位矩阵、H^T为观测矩阵H的转置形式。in,

is the optimal estimated state vector at time k,

is the predicted state vector at time k, P(k) is the updated value of the covariance matrix at time k,

is the predicted value of the covariance matrix at time k, K _G (k) is the Kalman matrix at time k,

is the prediction process error covariance matrix, σ _v is the process noise standard deviation of the velocity state transition equation, σ _a is the process noise standard deviation of the acceleration state transition equation,

is the sensor observation noise covariance matrix, σ _{v, m} is the standard deviation of the velocity sensor noise, σ _{a, m} is the standard deviation of the acceleration sensor noise, I is the identity matrix, and H ^T is the transposed form of the observation matrix H.

S214, performing moving average filtering to obtain a sample data set with longitudinal velocity and longitudinal acceleration of the vehicle;

In this technical solution, the data processing mainly includes state estimation of vehicle longitudinal velocity and longitudinal acceleration, estimation of rolling resistance acceleration term, and acceleration response delay time estimation for data alignment. The processed data forms a training set TrainingSet, the data input of the training set is [accelerator opening, brake opening, longitudinal speed of the vehicle, the current gear reduction ratio of the transmission], and the data label of the training set is [equivalent drive braking system] vehicle longitudinal acceleration];

State acquisition is based on GPS measurement of longitudinal velocity and IMU measurement of longitudinal acceleration, and the combination of limiting filtering, second-order linear Kalman filtering, and moving average filtering is used to estimate vehicle longitudinal velocity and longitudinal acceleration;

Perform moving average filtering to get the estimated results of vehicle longitudinal velocity and longitudinal acceleration that are smoother and closer to the real values.

In any of the above technical solutions, the S22 specifically includes: calibrating the vehicle coasting acceleration under the working conditions of zero accelerator opening and zero braking opening on the horizontal road surface, eliminating outlier data, and obtaining an estimated value of the rolling resistance acceleration term, according to The vehicle driving physical model and the vehicle braking physical model obtain the vehicle longitudinal acceleration of the equivalent driving braking system.

In this technical solution, the estimated value of the rolling resistance acceleration term is obtained by estimating the rolling resistance acceleration term, which is the gfcosθ term in the vehicle driving physical model and the vehicle braking physical model.

和

个采样数据的偏移对齐。In any of the above-mentioned technical solutions, the S23 specifically includes: calibrating the time of the accelerator command, the braking command and the time when the vehicle starts to accelerate and decelerate, count the delay times of acceleration and deceleration under different accelerators and brake openings, and calculate the average acceleration delay. The time τ _t and the average deceleration delay time τ _b are used to calculate the longitudinal acceleration of the vehicle with the equivalent driving and braking system.

and

Offset alignment of sample data.

In this technical solution, by calculating the delay time under different accelerator and brake opening degrees, and calculating the average acceleration delay time τ _t and average deceleration delay time τ _b , the sampling data can be offset and aligned, ensuring that The resulting data results are more realistic and smoother.

In any of the above technical solutions, the throttle neural network model adopts a multi-layer feedforward neural network, which specifically includes: an input layer, two hidden layers, and an output layer, and the two hidden layers respectively include 110 neurons and 20 neurons, the hidden layer uses Sigmoid as the activation function, the optimizer selects Adam, the loss function is the root mean square error, and the learning rate is 0.0001.

In this technical solution, the throttle neural network model adopts a multi-layer feedforward neural network, which can have better performance under large data samples.

In any of the above technical solutions, the braking neural network model adopts a multi-layer feedforward neural network, which specifically includes: an input layer, a hidden layer, and an output layer, the hidden layer includes 80 neurons, and the hidden layer includes 80 neurons. The layer uses tanh as the activation function, the optimizer selects Adam, the loss function is the root mean square error, and the learning rate is 0.0002.

In this technical solution, the braking neural network model adopts a multi-layer feedforward neural network, which can have better performance under large data samples.

制动神经网络模型：

车辆驱动物理模型为：

车辆制动物理模型为：

其中，

为驱动或制动系统产生的等效加速度，a_r为车辆实际加速度，

与a_r定义正值为加速，负值为减速；θ为道路坡度，定义正值为下坡道，负值为上坡道。；gfcosθ为滚动阻力加速度项，f为滚动阻力系数，g＝9.8/s²为重力加速度。In any of the above technical solutions, the accelerator neural network model:

Braking neural network model:

The vehicle drive physics model is:

The physical model of vehicle braking is:

in,

is the equivalent acceleration generated by the driving or braking system, a _r is the actual acceleration of the vehicle,

With a _r , positive value is defined as acceleration, negative value is deceleration; θ is road slope, positive value is defined as downhill, and negative value is defined as uphill. ; gfcosθ is the rolling resistance acceleration term, f is the rolling resistance coefficient, g=9.8/s ² is the gravitational acceleration.

Beneficial effects of the present invention:

1. Combine the vehicle driving physical model, the vehicle braking physical model with the accelerator neural network model and the braking neural network model to build a more accurate vehicle longitudinal dynamics model, which simplifies the construction of the overall vehicle longitudinal dynamics model as Only for the modeling problem of the equivalent acceleration generated by the drive system and the braking system, the obtained longitudinal dynamics model is more accurate.

2. The transmission ratio information of different gears of the gearbox is added to the accelerator neural network model for training, which can be applied to vehicles with few gearbox gears and the acceleration process is not smooth enough, such as mining wide-body dump trucks, muck trucks and other projects vehicle.

3. The variable of road gradient is added in the modeling, and the influence of road gradient on the actual acceleration performance of the vehicle is fully considered.

Additional aspects and advantages of embodiments in accordance with the invention will become apparent in the description section that follows, or learned through practice of embodiments in accordance with the invention.

Description of drawings

The drawings are for the purpose of illustrating specific embodiments only, and are not to be considered limiting of the present invention.

Fig. 1 is the overall flow chart of the method of the present invention;

Fig. 2 is the test diagram of automatic data collection of different accelerator/brake opening commands of the present invention;

Fig. 3 is the actual speed test result diagram of the vehicle under different accelerator/brake opening commands of the present invention;

Fig. 4 is the result diagram of vehicle test data speed state estimation result of the present invention;

Fig. 5 is the result diagram of the acceleration state estimation result of vehicle test data according to the present invention;

Fig. 6 is the throttle neural network model training result diagram of the present invention;

Fig. 7 is the training result diagram of the braking neural network model of the present invention;

8 is a three-dimensional diagram of the mapping relationship between vehicle speed, equivalent accelerator-braking acceleration and accelerator-braking opening degree for longitudinal speed control according to the present invention.

Detailed ways

In order to understand the above objects, features and advantages of the present invention more clearly, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments may be combined with each other in the case of no conflict.

Many specific details are set forth in the following description to facilitate a full understanding of the present invention. However, the present invention can also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited by the specific details disclosed below. Example limitations.

In the parameter symbols used in this embodiment, unless otherwise specified, the symbol marked with v in the lower corner represents the velocity direction correlation value of the parameter, the symbol marked with a in the lower corner represents the acceleration direction correlation value of the parameter, and the lower corner labelling The ones with X represent the correlation value of the parameter in the state vector, and the ones marked with ema in the lower corner represent the correlation value observed by the sensor of the parameter.

Referring to FIG. 1, a first aspect of the present invention provides a vehicle longitudinal dynamics identification method combining a neural network and a physical model, including: the longitudinal dynamics model is:

a _r =MODEL(v _x , φ _t , φ _b , _ig , θ _{road gradient} ), where: a _r is the actual acceleration of the vehicle, v _x is the longitudinal speed of the vehicle, φ _t is the accelerator opening, and φ _b is the braking Opening, _ig is the transmission reduction ratio, θ is the road gradient, and the longitudinal dynamics model is calibrated by the method, including the following steps: S1, drive the unmanned vehicle to drive, and record the driving data including the transmission gear; S2, according to the driving The data is processed to obtain a sample data set for training; S3, the throttle neural network model and the brake neural network model are trained respectively through the sample data set; S4, the trained throttle neural network model and brake neural network model are trained. Combined with the vehicle driving physical model and the vehicle braking physical model to obtain the longitudinal dynamics model of the unmanned vehicle; wherein, the driving data includes: accelerator opening, braking opening, longitudinal speed, longitudinal acceleration and transmission gear.

The invention provides a vehicle longitudinal dynamics identification method combined with a neural network and a physical model, aiming at the problem that the longitudinal speed control of an autonomous driving vehicle faces the complex longitudinal dynamics model, which has strong nonlinear and time-varying characteristics, which leads to difficult modeling. Combining the physical equation constructed by the vehicle's transmission gear and road gradient information with the neural network can better characterize the vehicle's longitudinal dynamics, and can be applied to more types of vehicles. The problem of constructing the overall longitudinal dynamics model of the vehicle is simplified as Only for the modeling problem of the equivalent acceleration generated by the drive system and the braking system, the obtained longitudinal dynamics model is more accurate.

Specifically, the vehicle physical model is derived based on the vehicle driving equation, including the vehicle driving physical model and the vehicle braking physical model; the vehicle driving equation is as follows:

F _t =F _f +F _i +F _w +F _j

为驱动力、F_f＝mgfcosθ为滚动阻力、F_i＝mgsinθ为坡度阻力、

为空气阻力，在低速情况下可以忽略、F_j＝kma_r为加速阻力、T_tq为发动机输出扭矩、r为车轮半径、i_g为变速箱传动比、i₀为主减速器传动比、η_T为传动系统效率、m为车辆质量、g＝9.8m/s²为重力加速度、θ为道路坡度、C_D为车辆空气阻力系数、A为迎风面积、v为车辆纵向速度、a_r为车辆纵向加速度、k为质量换算系数，f为滚动阻力系数。In the formula,

is the driving force, F _f = mgfcosθ is the rolling resistance, F _i = mgsinθ is the slope resistance,

is the air resistance, which can be ignored at low speed, F _j = _kmar is the _acceleration resistance, T _tq is the output torque of the engine, r is the wheel radius, ig is the transmission ratio of the gearbox, i ₀ is the transmission ratio of the main reducer, η _T is the efficiency of the transmission system, m is the mass of the vehicle, g=9.8m/s ² is the acceleration of gravity, θ is the road gradient, C _D is the air resistance coefficient of the vehicle, A is the windward area, v is the longitudinal speed of the vehicle, and a _r is the vehicle Longitudinal acceleration, k is the mass conversion coefficient, and f is the rolling resistance coefficient.

The vehicle drive physics model is:

The physical model of vehicle braking is:

为驱动或制动系统产生的等效加速度，a_r为车辆实际加速度，

与a_r定义正值为加速，负值为减速；θ为道路坡度，定义正值为下坡道，负值为上坡道。

is the equivalent acceleration generated by the driving or braking system, a _r is the actual acceleration of the vehicle,

With a _r , positive value is defined as acceleration, negative value is deceleration; θ is road slope, positive value is defined as downhill, and negative value is defined as uphill.

和制动神经网络模型

其中：v_x为车辆纵向速度、φ_t为油门开度、φ_b为制动开度、i_g为变速器减速比、

为等效驱动制动系统车辆纵向加速度。油门神经网络模型用于标定不同速度、不同油门开度所能够产生的等效加速度，制动神经网络模型用于标定不同速度、不同制动开度所能够产生的等效加速度，结合道路坡度加速度项和车辆滚动阻力加速度项，可以综合反映车辆在任何工况下的加速度表现，构成车辆纵向动力学模型。Specifically, the neural network model specifically includes a throttle neural network model

and braking neural network model

Where: v _x is the longitudinal speed of the vehicle, φ _t is the _accelerator opening, φ _b is the brake opening, ig is the transmission reduction ratio,

is the vehicle longitudinal acceleration for the equivalent drive braking system. The accelerator neural network model is used to calibrate the equivalent acceleration generated by different speeds and different accelerator openings, and the braking neural network model is used to calibrate the equivalent accelerations generated by different speeds and different braking openings, combined with the road gradient acceleration term and vehicle rolling resistance acceleration term, which can comprehensively reflect the acceleration performance of the vehicle under any working conditions, and constitute the vehicle longitudinal dynamics model.

In any of the above embodiments, through the reverse conversion of the longitudinal dynamics model, the mapping relationship of the vehicle longitudinal control parameters is obtained, which is specifically the following formula:

(φ _t , φ _b )=MAP(v _x , a _{desired speed} , θ _{road gradient} , i _g )

Wherein, the vehicle longitudinal control parameter mapping relationship is used for the control of the longitudinal speed of the unmanned vehicle, given the current vehicle operating state (v _x , θ _{road gradient} , i _g ) and the value of the longitudinal control rate (a _{desired speed} ) When , query to obtain the accelerator and brake opening values (φ _t , φ _b ) to be executed by the vehicle.

In this embodiment, the inverse dynamics model can be transformed into a vehicle control parameter mapping table and applied to the longitudinal speed control of the unmanned vehicle, so as to realize the desired acceleration output by the longitudinal speed control algorithm to the accelerator opening and braking that can be received by the executive layer. Mapping of opening instructions;

This mapping relationship can be used for vehicle longitudinal speed control. The mapping relationship of specific vehicle longitudinal control parameters means that the current vehicle transmission gear reduction ratio, vehicle longitudinal speed, road gradient, and expected acceleration are used as inputs to obtain the accelerator opening value and brake opening value that the vehicle control-by-wire system can receive. .

In any of the above-mentioned embodiments, the steps of S1 include: S11, setting different accelerator opening degrees and different brake opening degrees; S12, driving the unmanned vehicle on a level road and running at different speeds within a preset range; S13, collect the accelerator opening, brake opening, longitudinal speed, longitudinal acceleration, and transmission gear position when the unmanned vehicle is running; wherein, the preset range is 0km/h-30km/h, and the different accelerator openings and The brake opening is 0%, 10%, 20%, ..., 100% of the discrete opening command.

In this embodiment, a flat road is selected, and the automatic driving vehicle uses the pre-built acquisition program to conduct the acquisition experiment, and collects the vehicle longitudinal speed, longitudinal acceleration, and transmission gear (including the reduction ratio) according to the fixed accelerator/brake opening requirements. Wait for the vehicle status data to get real vehicle data [accelerator opening, brake opening, longitudinal speed, longitudinal acceleration, transmission gear];

It is mainly used in vehicles with a speed of less than 30km/h, such as mining wide-body dump trucks, muck trucks and other engineering vehicles, so the air resistance can be ignored.

In any of the above-mentioned embodiments, the step of S2 includes: S21, the longitudinal velocity and longitudinal acceleration of the vehicle are state estimated; S22, the rolling resistance acceleration term is estimated; S23, the acceleration response delay time is estimated, and data alignment is performed .

In this embodiment, the data is preprocessed to form a training data set for neural network training. The data processing mainly includes the following steps: estimating the longitudinal speed and longitudinal acceleration of the vehicle, estimating the rolling resistance acceleration term, Acceleration response delay time estimation.

In any of the above-mentioned embodiments, the estimation method of S21 is:

S211, firstly use the following formula to limit and filter the measured longitudinal acceleration, specifically:

Δa=a _mea (k)-a _mea (k-1)

Among them, Δa is the increase value from the (k-1)th sampling to the (k)th sampling, a _mea represents the measured value, a _{mea, limit} is the measured value after limiting;

S212, perform reprocessing to obtain the longitudinal velocity and longitudinal acceleration of the vehicle, specifically:

The discretized state transition equation and observation equation are constructed based on the vehicle kinematics model,

The state transition equation is:

即X(k)＝AX(k-1)+w(k-1)

That is, X(k)=AX(k-1)+w(k-1)

The observation equation is:

即Z(k)＝HX(k)+ε(k)

That is, Z(k)=HX(k)+ε(k)

为状态矢量、

为系统矩阵、

为传感器观测值、

为观测矩阵、T为数据采样周期、k为离散化的采样序列、

为系统的过程噪声、

为传感器观测噪声；in,

is the state vector,

is the system matrix,

are the sensor observations,

is the observation matrix, T is the data sampling period, k is the discretized sampling sequence,

is the process noise of the system,

Observe noise for the sensor;

S213, finally predict and update iteratively according to the following formula, specifically:

为k时刻的最优估计状态矢量、

为k时刻的预测状态矢量、P(k)为k时刻的协方差矩阵更新值、

为k时刻的协方差矩阵预测值、k_G(k)为k时刻的卡尔曼矩阵、

为预测过程误差协方差矩阵、σ_v为速度状态转移方程的过程噪声标准差、σ_a为加速度状态转移方程的过程噪声标准差、

为传感器观测噪声协方差矩阵、σ_v，m为速度传感器噪声的标准差、σ_a，m为加速度传感器噪声的标准差、I为单位矩阵、H^T为观测矩阵H的转置形式；in,

is the optimal estimated state vector at time k,

is the predicted state vector at time k, P(k) is the updated value of the covariance matrix at time k,

is the predicted value of the covariance matrix at time k, k _G (k) is the Kalman matrix at time k,

is the prediction process error covariance matrix, σ _v is the process noise standard deviation of the velocity state transition equation, σ _a is the process noise standard deviation of the acceleration state transition equation,

is the sensor observation noise covariance matrix, σ _{v, m} is the standard deviation of the speed sensor noise, σ _{a, m} is the standard deviation of the acceleration sensor noise, I is the identity matrix, H ^T is the transpose form of the observation matrix H;

S214 , performing moving average filtering to obtain a sample data set having the longitudinal velocity and longitudinal acceleration of the vehicle.

In this embodiment, data processing mainly includes state estimation of vehicle longitudinal speed and longitudinal acceleration, estimation of rolling resistance acceleration term, and acceleration response delay time estimation for data alignment. The processed data forms a training set TrainingSet, the data input of the training set is [accelerator opening, brake opening, longitudinal speed of the vehicle, the current gear reduction ratio of the transmission], and the data label of the training set is [equivalent drive braking system] Vehicle longitudinal acceleration], the speed sensor selects the GPS speed sensor, and the acceleration sensor selects the IMU acceleration sensor;

State acquisition is based on GPS measurement of longitudinal velocity and IMU measurement of longitudinal acceleration, and the combination of limiting filtering, second-order linear Kalman filtering, and moving average filtering is used to estimate vehicle longitudinal velocity and longitudinal acceleration;

Perform moving average filtering to obtain the estimated results of vehicle longitudinal speed and longitudinal acceleration that are smoother and closer to the real values;

为状态矢量，由每个采样时刻的车辆纵向速度和纵向加速度组成；

is the state vector, which consists of the longitudinal velocity and longitudinal acceleration of the vehicle at each sampling moment;

为传感器观测值，由每个采样时刻的GPS纵向测量速度和IMU纵向测量加速度组成；

is the sensor observation value, which consists of the GPS longitudinal measurement velocity and the IMU longitudinal measurement acceleration at each sampling time;

为系统的过程噪声，由纵向速度和纵向加速度进行状态转移预测时的两个过程噪声构成；

is the process noise of the system, which is composed of two process noises when the longitudinal velocity and longitudinal acceleration are used for state transition prediction;

加速度预测过程噪声服从高斯分布

均值为0，方差分别为

和

为传感器观测噪声，GPS速度传感器观测噪声服从高斯分布

IMU加速度传感器观测噪声服从高斯分布

The noise in the speed prediction process follows a Gaussian distribution

Acceleration prediction process noise obeys Gaussian distribution

The mean is 0 and the variances are

and

For the sensor observation noise, the GPS speed sensor observation noise obeys a Gaussian distribution

The observation noise of the IMU accelerometer obeys a Gaussian distribution

In any of the above-mentioned embodiments, S22 specifically includes: calibrating the vehicle coasting acceleration under the operating conditions of zero accelerator opening and zero braking opening on the horizontal road surface, eliminating outlier data, obtaining an estimated value of the rolling resistance acceleration term, and driving the vehicle according to the vehicle drive. The physical model and the vehicle braking physics model obtain the vehicle longitudinal acceleration of the equivalent drive braking system.

In this embodiment, the estimated value of the rolling resistance acceleration term is obtained by estimating the rolling resistance acceleration term, which is the gfcosθ term in the vehicle driving physical model and the vehicle braking physical model.

和

个采样数据的偏移对齐。In any of the above-mentioned embodiments, S23 specifically includes: calibrating the accelerator command, the braking command time and the time when the vehicle starts to accelerate and decelerate, count the delay times of acceleration and deceleration under different accelerators and brake openings, and calculate the average acceleration delay time τ. _t and the average deceleration delay time τ _b , the longitudinal acceleration of the vehicle in the equivalent driving braking system is calculated.

and

Offset alignment of sample data.

In this embodiment, by counting the delay times under different accelerator and brake opening degrees, and calculating the average acceleration delay time τ _t and average deceleration delay time τ _b , the sample data can be offset and aligned to ensure that The resulting data results are more realistic and smoother.

In any of the above embodiments, the throttle neural network model adopts a multi-layer feedforward neural network, which specifically includes: an input layer, two hidden layers, and an output layer, and the two hidden layers respectively include 110 neurons and 20 neurons. , the hidden layer uses Sigmoid as the activation function, the optimizer selects Adam, the loss function is the root mean square error, and the learning rate is 0.0001.

In this embodiment, the throttle neural network model adopts a multi-layer feedforward neural network, which can have better performance under large data samples.

In any of the above embodiments, the braking neural network model adopts a multi-layer feedforward neural network, which specifically includes: an input layer, a hidden layer, and an output layer, the hidden layer includes 80 neurons, and the hidden layer uses tanh as the activation function. , the optimizer chooses Adam, the loss function is the root mean square error, and the learning rate is 0.0002.

In this embodiment, the braking neural network model adopts a multi-layer feedforward neural network, which can have better performance under large data samples.

制动神经网络模型：

车辆驱动物理模型为：

车辆制动物理模型为：

其中，

为驱动或制动系统产生的等效加速度，a_r为车辆实际加速度，

与a_r定义正值为加速，负值为减速；θ为道路坡度，定义正值为下坡道，负值为上坡道。；gfcosθ为滚动阻力加速度项，f为滚动阻力系数，g＝9.8m/s²为重力加速度。In any of the above embodiments, the throttle neural network model:

Braking neural network model:

The vehicle drive physics model is:

The physical model of vehicle braking is:

in,

is the equivalent acceleration generated by the driving or braking system, a _r is the actual acceleration of the vehicle,

With a _r , positive value is defined as acceleration, negative value is deceleration; θ is road slope, positive value is defined as downhill, and negative value is defined as uphill. ; gfcosθ is the rolling resistance acceleration term, f is the rolling resistance coefficient, g=9.8m/s ² is the gravitational acceleration.

Example 1

As shown in Figure 2-8, the vehicle is a mining wide-body dump truck with a maximum speed of 40km/h, and the experimental data collection site is a flat hard dirt road.

Step 1: Start the programmed automatic driving program, run the unmanned mining wide-body dump truck on the level road, and set different accelerator opening degrees (0.0, 0.1, 0.2, ..., 0.9, 1.0) and different brake opening degrees in the program. (0.0, 0.1, 0.2, …, 0.9, 1.0) to collect driving data at different speeds. The vehicle speed is accelerated from 0km/h to the maximum speed of 23km/h and then decelerated to 0km/h. The data collection test is shown in Figure 2 and Figure 3.

Step 2: Preprocess the data to form a training data set for neural network training. The data processing mainly includes the following steps: estimating the longitudinal velocity and longitudinal acceleration of the vehicle, estimating the rolling resistance acceleration term, and estimating the acceleration response delay time.

State estimation for vehicle longitudinal velocity and longitudinal acceleration: including limiting filtering, second-order linear Kalman filtering, and moving average filtering. The results of velocity and acceleration estimation from the data collected in the final experiment are shown in Figures 4 and 5 below.

The third step: carry out network training, conduct network training on the sample data set, the data is collected from the horizontal road, the equivalent driving and braking system vehicle longitudinal acceleration is divided by the transmission speed of the reducer at each sampling time, and the corrected equivalent driving The vehicle longitudinal acceleration of the braking system is used to train the accelerator neural network model and the braking neural network model.

The fourth step, model generation, the results of the training of the accelerator neural network model and the braking neural network model are shown in Figures 6 and 7 below. Combined with the vehicle driving physical model and the vehicle braking physical model, the precise dynamics of a specific vehicle is finally obtained. Model a _r =MODEL(v _x , φ _t , φ _b , _ig , θ _{road gradient} ).

The precise dynamic model of the mining wide-body dump truck can be reversely converted to obtain the longitudinal control parameter mapping relationship of the mining wide-body dump truck (φ _t , φ _b )=MAP(v _x , a _{desired speed} , θ _{road gradient} , i _g ), which can be applied to the longitudinal speed control of unmanned vehicles to achieve a reasonable mapping of the expected acceleration to the accelerator and brake opening in the speed control algorithm. As shown in Figure 8, under the condition of horizontal road surface, the speed of the mining wide-body dump truck, the equivalent accelerator braking acceleration and the accelerator braking opening (-1 to 0 are the specified opening, 0 to 1 is the accelerator opening In the process of use, the expected acceleration can be converted into the equivalent accelerator braking acceleration in combination with the real-time gear position and road gradient information, and then check the mapping relationship to obtain the accelerator braking opening, which is used for the vehicle's Longitudinal speed control

In the description of the present invention, it should be understood that the terms "portrait", "horizontal", "upper", "lower", "front", "rear", "left", "right", "vertical", The orientation or positional relationship indicated by "horizontal", "top", "bottom", "inner", "outer", etc. is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present invention, rather than indicating or It is implied that the device or element referred to must have a particular orientation, be constructed and operate in a particular orientation, and therefore should not be construed as limiting the invention.

The above embodiments are only to describe the preferred mode of the present invention, but not to limit the scope of the present invention. On the premise of not departing from the design spirit of the present invention, those of ordinary skill in the art can make various deformations and modifications to the technical solutions of the present invention. Improvements should all fall within the protection scope determined by the claims of the present invention.

Claims (9)

1. A vehicle longitudinal dynamics identification method combining a neural network and a physical model is characterized in that the longitudinal dynamics model is as follows:

ar＝MODEL(v _x ，φ _t ，φ _b ，i _g ，θ _{road grade} ) Wherein: a is _r As actual acceleration, v, of the vehicle _x Is the longitudinal speed, phi, of the vehicle _t Is the throttle opening, phi _b Is the brake opening degree i _g For transmission reduction ratio, θ being road grade, said longitudinal dynamics model is calibrated by said method, comprising the steps of:

s1, driving an unmanned vehicle to run, and recording running data;

s2, processing is carried out according to the driving data to obtain a sample data set for training;

s3, respectively training the accelerator neural network model and the brake neural network model through the sample data set;

s4, combining the trained accelerator neural network model and the trained braking neural network model with the vehicle driving physical model and the vehicle braking physical model to obtain a longitudinal dynamic model of the unmanned vehicle;

wherein the travel data includes: throttle opening, brake opening, longitudinal speed, longitudinal acceleration, and transmission gear.

2. The vehicle longitudinal dynamics identification method combining the neural network and the physical model according to claim 1, wherein a vehicle longitudinal control parameter mapping relationship is obtained through inverse transformation of the longitudinal dynamics model, and specifically is the following formula:

(φ _t ，φ _b )＝MAP(v _x ，a _{desired speed} ，θ _{Road grade} ，i _g )

Wherein the vehicle longitudinal control parameter mapping is used for control of the longitudinal speed of the unmanned vehicle given the current vehicle operating state (v) _x ，θ _{Road grade} ，i _g ) And longitudinal control rate value (a) _{Desired speed} ) Then, the opening value (phi) of the accelerator and the brake to be executed by the vehicle is obtained through inquiry _t ，φ _b )。

3. The method for identifying longitudinal dynamics of a vehicle combining a neural network and a physical model according to claim 1, wherein the step of S1 comprises:

s11, setting different accelerator opening degrees and different brake opening degrees;

s12, driving the unmanned vehicle to run on a horizontal road at different speeds within a preset range;

s13, collecting the opening degree of an accelerator, the opening degree of a brake, the longitudinal speed, the longitudinal acceleration and the gear position of a transmission when the unmanned vehicle runs;

the preset range is 0km/h-30km/h, and the different accelerator opening degrees and brake opening degrees are 0%,10%,20%, … … and 100% of discretization opening degree instructions.

4. The method for identifying longitudinal dynamics of a vehicle combining a neural network and a physical model according to claim 1, wherein the step of S2 comprises:

s21, carrying out state estimation on the longitudinal speed and the longitudinal acceleration of the vehicle;

s22, estimating a rolling resistance acceleration item;

and S23, estimating the acceleration response delay time and aligning data.

5. The method for identifying longitudinal dynamics of a vehicle combining a neural network and a physical model according to claim 4, wherein the estimation method of S21 is as follows:

s211, performing amplitude limiting filtering processing on the measured longitudinal acceleration by using the following formula, specifically:

Δ _a ＝a _mea (k)-a _mea (k-1)

where Δ a is the added value from sample (k-1) to sample (k), a _mea Represents the measured value, a _mea，limit Is the measured value after amplitude limiting;

s212, reprocessing is carried out to obtain the longitudinal speed and the longitudinal acceleration of the vehicle, and the concrete steps are as follows:

a discretization state transition equation and an observation equation are constructed based on the vehicle kinematics model,

the state transition equation is:

i.e., X (k) = AX (k-1) + w (k-1)

The observation equation is:

that is, Z (k) = HX (k) + ε (k)

Wherein,

is a state vector,

Is a system matrix,

For the observed value of the sensor,

Is an observation matrix, T is a data sampling period, k is a discretized sampling sequence,

Is the process noise of the system,

Observing noise for a sensor, wherein T is a data sampling period and k is a discretized sampling sequence;

s213, continuously iterating to predict and update according to the following formula:

wherein,

for the best estimated state vector at time k,

For the predicted state vector at time k, P (k) is the covariance matrix update value at time k,

As covariance matrix prediction value at time K, K _G (k) Is a Kalman matrix at time k,

For predicting the process error covariance matrix, sigma _v Process noise standard deviation, σ, for velocity state transition equation _a Is the process noise standard deviation of the acceleration state transition equation,

Observation of the noise covariance matrix, σ, for the sensor _v，m Is the standard deviation, sigma, of the velocity sensor noise _a，m Is the standard deviation of noise of the acceleration sensor, I is an identity matrix, H ^T Is the transposed form of the observation matrix H;

and S214, performing moving average filtering to obtain a sample data set with the longitudinal speed and the longitudinal acceleration of the vehicle.

6. The method for identifying longitudinal dynamics of a vehicle combining a neural network and a physical model according to claim 4, wherein the step S22 specifically comprises:

the method comprises the steps of calibrating vehicle sliding acceleration under the working conditions of zero accelerator opening and zero brake opening of a horizontal road surface, eliminating outlier data to obtain an estimated value of a rolling resistance acceleration item, and obtaining the vehicle longitudinal acceleration of an equivalent driving brake system according to a vehicle driving physical model and a vehicle braking physical model.

7. The method for identifying longitudinal dynamics of a vehicle combining a neural network and a physical model according to claim 4, wherein the step S23 specifically comprises:

calibrating the accelerator instruction and brake instruction time and the time when the vehicle starts to accelerate and decelerate, counting the delay time of acceleration and deceleration under different accelerator and brake opening degrees, and calculating the average acceleration delay time tau _t And average deceleration delay time τ _b To the longitudinal acceleration of the vehicle in the equivalent drive brake system

And

the offsets of the individual sample data are aligned.

8. The method for identifying longitudinal dynamics of a vehicle by combining a neural network and a physical model according to claim 1, wherein the throttle neural network model adopts a multilayer feedforward neural network, and specifically comprises: the neural network comprises an input layer, two hidden layers and an output layer, wherein the two hidden layers respectively comprise 110 neurons and 20 neurons, the hidden layers adopt Sigmoid as an activation function, an optimizer selects Adam, a loss function is root-mean-square error, and the learning rate is 0.0001; and/or

The braking neural network model adopts a multilayer feedforward neural network, and specifically comprises the following steps: the neural network comprises an input layer, a hidden layer and an output layer, wherein the hidden layer comprises 80 neurons, the hidden layer adopts tanh as an activation function, the optimizer selects Adam, a loss function is root mean square error, and a learning rate is 0.0002.

9. The method of claim 1, wherein the neural network is used to identify the longitudinal dynamics of the vehicle,

accelerator neural network model:

braking neural network model:

the vehicle driving physical model is as follows:

the physical model of the vehicle brake is as follows:

wherein,

equivalent acceleration for driving or braking systems, a _r Is the actual acceleration of the vehicle,

and a _r Defining positive values as acceleration and negative values as deceleration; theta is the road gradient and defines positive values as downhill slopes and negative values as uphill slopes. (ii) a gfcos theta is a rolling resistance acceleration term, f is a rolling resistance coefficient, g =9.8m/s ² Is the acceleration of gravity.

Images